高频稀疏偏置先验的全色锐化技术  

High-frequency Sparse Bias Prior with Pansharpening

作　　者：许书源 潘晓航 XU Shuyuan;PAN Xiaohang(School of Communication,Soochow University,Suzhou 215129,Jiangsu,China)

机构地区：[1]苏州大学传媒学院,江苏苏州215129

出　　处：《实验室研究与探索》2024年第8期18-23,51,共7页Research and Exploration In Laboratory

基　　金：江苏省高校实验室研究会研究课题项目(GS2022BZZ37)。

摘　　要：针对全色锐化任务中全色图像和多光谱图像之间的不对齐现象,提出了一种观测模型,通过确保全色图像和多光谱图像高频成分的一致性来更好地描述光谱退化。此外,还设计了一个稀疏偏置先验,以进一步减少因离群点而造成的建模误差。在此基础上,构建了一个用于全色锐化的目标函数,并将迭代求解过程展开为深度卷积网络,其中空间退化和光谱退化过程都是通过设计特定的子网络进行逼近。在降分辨率和全分辨率数据集上的实验结果均验证了所提方法的有效性。In pansharpening tasks,misalignment is inevitable due to the fact that panchromatic(PAN)images and multispectral(MS)images are collected by different sensors.However,most existing pancsharpening methods completely ignore this inherent misalignment phenomenon.To address this issue,a novel observational model is proposed that better describes spectral degradation by ensuring consistency between the high-frequency components of PAN and HRMS images.Additionally,a sparse bias prior is designed to further reduce modeling errors caused by outliers.Based on these improvements,a new objective function for pansharpening is constructed,and the iterative solution is unrolled into a deep convolutional network,where spatial and spectral degradations are approximated using specific subnets.Experimental results at reduced-resolution and full-resolution are reported to demonstrate the effectiveness of the proposed method.

关 键 词：遥感图像融合 全色锐化 深度学习 深度展开 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]